Learning to Extract Entities from Labeled and Unlabeled Text

We describe and evaluate algorithms for learning to extract semantic classes from sentences in text documents, using the minimum of training information. The thesis of this research is that we can efficiently automate information extraction, that is, learn from tens of examples of labeled training data instead of requiring thousands, by exploiting redundancy and separability of the features noun-phrases and contexts. We exploit this redundancy and separability in two ways: (1) in the algorithms for learning semantic classes, and (2) in novel algorithms for active learning, leading to better extractors for a given amount of user labeling effort.